Converge
Open the app

Week 14 · Season 4 · Will

A clean desk

You can keep an agent's context small and sharp, give it a long-term memory that forgets on purpose, gate risky actions behind a person's yes, and grade the whole road it takes.

Open week 14 in the app

This week

  • Context rot and the attention budget
  • Squeeze, fetch, delegate: compaction, pointers, sub-agents
  • Long-term memory: events, facts, habits
  • Human in the loop: go, ask, never
  • Trajectory evals

How you learn

Before you write any memory code, measure context rot yourself: plant a note at five depths, ask twenty times each, and chart the recall.

What you build

FRIDAY gets a long-term memory (events, facts, habits) that survives a restart and forgets on purpose, plus an approval gate for risky actions, both proved with trajectory tests.

Stretch: Add compaction to FRIDAY's long chats: squeeze at a threshold, keep pinned rules and tool pairs, and test that a planted rule survives three squeezes.

What you ship

Project 4: The morning briefing. Every morning FRIDAY gathers your calendar, tasks and notes and briefs you, by itself, within budget.

  • MEMORY.md, GATE.md and a diagram of the cabinet and the gate
  • memory.py and gate.py with trajectory tests that pass in CI

Lessons

  1. Context rot

    After this deck you can explain context rot in plain words, say why a note in the middle of a long context gets missed, and measure it with a plant-a-note experiment.

    11 minFull access

  2. Squeeze, fetch, delegate

    After this deck you can keep a long-running agent's context small with compaction, fetching on demand and sub-agents, and choose the right habit for a job on a budget.

    12 minFull access

  3. Memory that forgets on purpose

    After this deck you can give FRIDAY a long-term memory with three drawers (what happened, what's true, how you like things), and decide what gets stored, what gets fetched and what gets forgotten.

    12 minFull access

  4. Ask before you act

    After this deck you can sort FRIDAY's tools into go, ask and never, design an approval card a person will actually read, and pause a run safely until they answer.

    12 minFull access

  5. Grade the path

    After this deck you can write a trajectory eval that checks the tool FRIDAY used first, its number of steps, forbidden actions and cost, not just its final answer.

    12 minFull access

  6. Go deeper: context engineering

    After this deck you can compact a long context safely (a threshold, recent turns kept word for word, tool request and result pairs kept together, notes kept off the desk), and compare memory stores and how they rank what to fetch.

    13 minFull accessGo deeper · optional

Labs

Python in the browser, next to the lessons.

Weekend build

Skills you prove

Context
Design the context for a long-running agent: compaction, memory, sub-agents

Hand-picked official docs, papers and talks.